I can avoid the ko problem by storing the depth of each node and
never creating a link from a node at depth d except to a node at
depth d+1. This prevents any cycles, at the (minimal, in my opinion)
cost of omitting transpositions of different lengths.
I need the move for two reasons:
1) In performing UCT, I need to traverse my children, find the value
and confidence bound for each one, and then choose the move leading
to the "best" one. This requires knowing which move leads to which
child node.
2) In testing, I like to be able to print out the tree.
Peter Drake
Assistant Professor of Computer Science
Lewis & Clark College
http://www.lclark.edu/~drake/
On Dec 7, 2006, at 2:58 PM, House, Jason J. wrote:
DAG's have a problem with graph history, especially with super ko
considerations. Do you need the parent play for more than ko
considerations?
-----Original Message-----
From: [EMAIL PROTECTED]
[mailto:[EMAIL PROTECTED] On Behalf Of Peter Drake
Sent: Thursday, December 07, 2006 5:47 PM
To: Computer Go
Subject: [computer-go] firstChild/nextSibling in a DAG
(This is all within the context of Monte Carlo.)
Is anyone storing a search DAG (as opposed to a tree) and using the
firstChild/nextSibling representation? I'm having trouble seeing how
this would work, since when you traverse children (e.g., in UCT) you
have to know which move is associated with which child node. If a
node might have more than one parent, the node can't store its last
move.
Any clever solutions?
If not, any opinions (or better yet, evidence) as to whether the
space savings or the DAG transposition table is more valuable?
Peter Drake
Assistant Professor of Computer Science
Lewis & Clark College
http://www.lclark.edu/~drake/
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
